Embracing Model Uncertainty: Strategies for Response Pooling and Model Averaging

Econometricians modeling Stated Preference (SP) data, and most other types of data, are confronted with the uncomfortable reality that our knowledge of the “true” model is limited, with only certain variables suggested by the application at hand and general classes of functional forms and error structures suggested by the literature. Accepting our limited knowledge, we pursue strategies for analyzing SP data that are more robust to uncertainties in our knowledge of the true model. These include non-parametric and parametric likelihood-based tests of pooling responses from different elicitation formats, and a frequentist-based model averaging approach for estimating willingness to pay functions. We argue that these strategies lead to increased econometric integrity and empower the ultimate users of models, such as policy decision-makers and even juries, to better assess the robustness of the results. We apply these approaches to an SP survey of saltwater angling in Alaska which utilized split-sample rankings and ratings elicitation methods. While an important goal of our paper is to develop practicable modeling strategies that will ultimately lead to more robust conclusions and more confidence by the users of SP results, an equally important goal is to engender a critical discussion of how we can make the analysis of SP data more robust.